Identification of Mechanical Properties of Thin-Film Elastoplastic Materials by Machine Learning

Nanoindentation can effectively evaluate the mechanical properties of materials in the form of bulk and coating. However, the relationship between the indentation response and the stress–strain curve of thin-film elastoplastic materials is complex and thus difficult to be elucidated using traditional physics-based, empirical or statistical models. In this study, the convolutional neural network (CNN), as a practical machine learning method, is adopted and trained to rapidly obtain the mechanical properties of thin-film elastoplastic materials using nanoindentation. The proposed method is targeted for efficiently predicting mechanical properties of thin-film materials from the applied load–penetration depth curve. Combined with the power-law model to describe the elastoplastic characteristics, a dataset comprising 228 nanoindentation cases with wide ranges of material properties is numerically simulated by ABAQUS and the corresponding results are adopted for the CNN training and validating. By addressing the important elastoplastic properties characterized by elastic modulus, yield strength, and hardening exponent, the impacts of CNN’s architecture and training epochs on the predicting performance are investigated in detail. By varying the number of convolutional layers, the influence of mechanical parameters of thin-film materials on the CNN prediction accuracy is discussed. The results show that compared with the traditional reverse algorithm, CNN can greatly reduce the computational complexity and computation time and has better prediction accuracy for the constitutive parameters of thin-film elastoplastic materials.